Evaluating the performance of machine learning methods and Geographic Information System (GIS) in identifying groundwater potential zones in Tehran-Karaj plain, Iran

Abstract

Approaches based on Geographic Information Systems (GISs) are the most common and influential methods in identifying and mapping Groundwater Potential Zones (GWPZ). Moreover, employing Machine Learning (ML) algorithms in modeling the influencing parameters on GWPZ can improve the performance of these models. Therefore, the present study aims to assess the efficiency of ML methods, including Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), and Extreme Gradient Boost (XGB) models, along with influencing factors on GWPZ in GIS environment. Hence, the Tehran-Karaj plain in Iran, with approximately 5156 km2 area and relatively diverse elevations (818 to 4374 m), including flat and mountainous areas, was selected as the study area. Information from 105 wells was collected and divided into 75% for training and 25% for validation of the models. The effects of twelve parameters were investigated in three categories: topography, hydrology, and vegetation/geology (four parameters for each item). Moreover, LR, RF, SVM, and XGB models were applied to calculate the scores of each influential factor and create maps of GWPZ for the study area. Results of assessing the accuracy of GWPZ methods using the Receiver Operating Characteristic curve (ROC) revealed that the highest Area Under the Curve (AUC) value (89%) was obtained for the XGB model compared to RF, LR, and SVM models (with values of 87.4, 76, and 75.3%, respectively). Moreover, XGB and RF show higher Kappa coefficient values, with the values of 80% and 65%, respectively. This indicated the tree-based models' high performance (i.e., XGB and RF) compared to other models. Model outputs also demonstrated that topographic parameters (i.e., altitude levels and slope) strongly influenced the GWPZ predictions of the used models in the study area. In general, results indicated the high performance of ML algorithms and GIS for identifying GWPZ.